brain-of-reese/tests/unit/test_agent.py

"""Unit: the grounded-turn agent loop (phase 37, ``app.rag.agent``; the
harness-aligned ``ls``/``read``/``grep`` surface, phase 70).

A scripted fake LLM (canned stream sequences) + monkeypatched
``list_catalog`` / ``list_source_names`` / ``find_document`` /
``all_documents`` — no database, no network. Covers the loop mechanics:
the ls → read (combined ``source/path``) → answer happy path (event
order, holder state, the tools staying offered on every request —
phase 45 removed the per-tool budgets, the assistant/tool message
history), the ``ls`` scoping (no-arg full catalog in the phase-63
labeled-field format, a one-source scope, a known source with 0
documents → ``0 documents:`` counted, an unknown-source refusal that
counts nothing), ``read`` on the canonical combined form (split at the
FIRST slash, full content, the bare-source-name refusal, the
already-in-context dedupe, missing-args refusals), the phase-68 ``grep``
contract under its new name (the locked A5 pins: fixed substring,
case-insensitive, 20-cap in catalog order, 200-char lines, locator-only
— ``read_docs`` untouched, no-match lines counted), the round cap
forcing a final no-tools answer, the kill switch
(``agent_max_rounds=0`` single-call path), and the phase-67 per-round
retries (a dead-then-recovered round restarts before its first piece
with a ``RetryPiece``; a mid-stream drop stays terminal — locked A2;
the forced final no-tools call retries too; ``llm_retries=0`` is one
plain attempt; retries are invisible to the round cap; consumer abandon
mid-retry-sleep leaks nothing).
"""
from __future__ import annotations

import asyncio
import json
import logging
import uuid
from collections.abc import AsyncGenerator, AsyncIterator
from copy import deepcopy
from typing import TYPE_CHECKING, Any, cast

import pytest
from sqlalchemy.orm import Session

from app.config import Settings
from app.models import Document, GitSource
from app.rag import agent
from app.rag.agent import (
    AGENT_TOOLS,
    AgentHolder,
    MalformedReplyError,
    run_agent,
)
from app.rag.llm import LLMClient, LLMError, RetryPiece, StreamPiece, ToolCallPiece
from app.rag.prompts import TOOLS_SECTION, _base, build_deflect_prompt, build_high_prompt

if TYPE_CHECKING:
    from app.rag.scaffolding import ScaffoldingFilter


def _settings(**kwargs: Any) -> Settings:
    kwargs.setdefault("_env_file", None)
    return Settings(**kwargs)  # pyright: ignore[reportCallIssue]


def _doc(source: str, path: str, title: str = "Title", content: str = "CONTENT") -> Document:
    return Document(
        id=uuid.uuid4(),
        source=source,
        path=path,
        full_path=f"/tmp/{path}",
        title=title,
        content=content,
        content_hash="0" * 64,
    )


class ScriptedLLM:
    """Canned stream sequences; records every ``chat_stream`` request so
    the tests can assert on the messages and the ``tools`` passthrough.
    Phase 71: when the caller passes a ``ScaffoldingFilter``, the canned
    content pieces are fed through it exactly like
    ``LLMClient.chat_stream`` (an empty clean result yields nothing; the
    held tail is flushed on normal completion) — so a scaffolding-only
    canned round streams no content pieces and leaves ``stripped_chars``
    behind for the recovery policy to key on."""

    def __init__(self, *streams: list[StreamPiece | ToolCallPiece]) -> None:
        self.streams: list[list[StreamPiece | ToolCallPiece]] = list(streams)
        self.requests: list[tuple[list[dict[str, Any]], list[dict[str, Any]] | None]] = []

    async def chat_stream(
        self,
        messages: list[dict[str, str]],
        tools: list[dict[str, Any]] | None = None,
        scaffolding: ScaffoldingFilter | None = None,
    ) -> AsyncIterator[StreamPiece | ToolCallPiece]:
        self.requests.append((deepcopy(messages), tools))
        if not self.streams:
            raise AssertionError("ScriptedLLM ran out of canned streams")
        pieces = self.streams.pop(0)
        if scaffolding is None:
            for piece in pieces:
                yield piece
            return
        for piece in pieces:
            if isinstance(piece, StreamPiece) and piece.kind == "content":
                cleaned = scaffolding.feed(piece.text)
                if cleaned:
                    yield StreamPiece("content", cleaned)
            else:
                yield piece
        tail = scaffolding.flush()
        if tail:
            yield StreamPiece("content", tail)


async def _run(
    llm: ScriptedLLM | FailingLLM,
    holder: AgentHolder,
    settings: Settings,
    seed_docs: list[Document] | None = None,
) -> list[StreamPiece | ToolCallPiece | RetryPiece]:
    out: list[StreamPiece | ToolCallPiece | RetryPiece] = []
    async for piece in run_agent(
        cast("LLMClient", llm),
        cast("Session", None),
        system_prompt="SYSTEM_PROMPT",
        user_message="QUESTION",
        seed_docs=seed_docs or [],
        settings=settings,
        holder=holder,
    ):
        out.append(piece)
    return out


# ---------- AGENT_TOOLS shape (phase 70: ls / read / grep) ----------


def test_agent_tools_names_and_parameters() -> None:
    by_name = {t["function"]["name"]: t for t in AGENT_TOOLS}
    assert len(AGENT_TOOLS) == 3  # ls / read / grep (phase 70)
    assert set(by_name) == {"ls", "read", "grep"}
    # The phase-37/68 names exist nowhere in the tool surface.
    assert not set(by_name) & {"list_documents", "read_document", "search_documents"}
    assert all(t["type"] == "function" for t in AGENT_TOOLS)
    ls = by_name["ls"]["function"]
    assert ls["description"] == (
        "List the indexed documents as `source: X | path: Y | title: Z` lines."
    )
    ls_params = ls["parameters"]
    assert ls_params["type"] == "object"
    assert ls_params["required"] == []  # path is optional
    assert set(ls_params["properties"]) == {"path"}
    assert ls_params["properties"]["path"]["type"] == "string"
    assert ls_params["properties"]["path"]["description"] == (
        "Source name to list one source's documents (e.g. 'homelab'); "
        "omit to list every document."
    )
    read = by_name["read"]["function"]
    assert read["description"] == (
        "Add the full content of one indexed document to your context."
    )
    read_params = read["parameters"]
    assert read_params["type"] == "object"
    assert read_params["required"] == ["path"]
    assert set(read_params["properties"]) == {"path"}
    assert read_params["properties"]["path"]["type"] == "string"
    # The combined source/path string is the canonical document identity
    # (phase 70) — the description pins it with a worked example.
    assert read_params["properties"]["path"]["description"] == (
        "The document to add to your context, as the combined "
        "`source/path` string exactly as shown in the `ls` output (e.g. "
        "'homelab/active/container_caddy/caddy.md')."
    )
    grep = by_name["grep"]["function"]
    assert grep["description"] == (
        "Search the indexed documents for an exact string "
        "(case-insensitive) and return up to 20 matching lines as "
        "`source/path:line: text` — a locator, not a context-adder: "
        "read the winner with `read`."
    )
    grep_params = grep["parameters"]
    assert grep_params["type"] == "object"
    assert grep_params["required"] == ["pattern"]
    assert set(grep_params["properties"]) == {"pattern", "path"}
    assert all(p["type"] == "string" for p in grep_params["properties"].values())
    assert grep_params["properties"]["pattern"]["description"] == (
        "The exact text to search for (a plain substring, not a regex)"
    )
    assert grep_params["properties"]["path"]["description"] == (
        "Limit the search to one document, as a combined `source/path` "
        "string from the `ls` output (omit to search every document)."
    )


def test_agent_tools_order_is_ls_read_grep() -> None:
    """The listing → context → locator order the prompt teaches (the API
    layer and the mock key off the names)."""
    assert [t["function"]["name"] for t in AGENT_TOOLS] == ["ls", "read", "grep"]


def test_refusal_constants_are_harness_aligned() -> None:
    """The updated module-level refusal lines (the names moved to the
    harness surface; ALREADY_IN_CONTEXT / UNKNOWN_TOOL unchanged)."""
    assert agent.ALREADY_IN_CONTEXT == "Already in your context."
    assert agent.UNKNOWN_TOOL == "Unknown tool."
    assert agent.MISSING_READ_ARGS == "read requires a string argument 'path'."
    assert agent.MISSING_SEARCH_ARGS == "grep requires a string argument 'pattern'."


# ---------- list_source_names (the scoped ls registry join) ----------


def test_list_source_names_resolves_registry_rows(
    monkeypatch: pytest.MonkeyPatch,
) -> None:
    """Names resolve exactly as the import pipeline indexes them (the
    phase-69 ``resolve_source_name`` expressions — reuse, not
    re-derivation), deduped (two rows resolving to the same name share
    documents), in registry order."""
    rows = [
        GitSource(url="https://github.com/reese/homelab.git", kind="git"),
        GitSource(
            url="/srv/reese/deployments", kind="local", path="/srv/reese/deployments"
        ),
        # The phase-69 sibling case: a second row, same resolved name.
        GitSource(url="https://github.com/reese/homelab", kind="git"),
    ]
    monkeypatch.setattr(agent, "effective_sources", lambda db: (rows, "db"))
    assert agent.list_source_names(cast("Session", object())) == [
        "homelab",
        "deployments",
    ]


def test_list_source_names_empty_registry(monkeypatch: pytest.MonkeyPatch) -> None:
    monkeypatch.setattr(agent, "effective_sources", lambda db: ([], "env"))
    assert agent.list_source_names(cast("Session", object())) == []


# ---------- happy path: ls → read (combined path) → answer ----------


def test_ls_then_read_then_answer(monkeypatch: pytest.MonkeyPatch) -> None:
    catalog = [
        ("Deployments", "backups.md", "Backup Strategy"),
        ("Homelab", "aws-route53.md", "AWS Route53 Records"),
    ]
    monkeypatch.setattr(agent, "list_catalog", lambda db: catalog)
    target = _doc("Homelab", "aws-route53.md", "AWS Route53 Records", "R53-CONTENT")
    calls: list[tuple[str, str]] = []

    def _find(db: Any, source: str, path: str) -> Document | None:
        calls.append((source, path))
        return target if (source, path) == ("Homelab", "aws-route53.md") else None

    monkeypatch.setattr(agent, "find_document", _find)
    seed = [_doc("Homelab", "kubernetes.md", "Kubernetes", "K8S-CONTENT")]
    holder = AgentHolder()
    llm = ScriptedLLM(
        [ToolCallPiece(id="call_1", name="ls", arguments={})],
        [
            ToolCallPiece(
                id="call_2",
                name="read",
                arguments={"path": "Homelab/aws-route53.md"},
            )
        ],
        [StreamPiece("thinking", "hmm "), StreamPiece("content", "Done! ")],
    )

    pieces = asyncio.run(_run(llm, holder, _settings(), seed_docs=seed))

    # Event order: tool pieces before the answer content/thinking.
    assert [type(p) for p in pieces] == [
        ToolCallPiece,
        ToolCallPiece,
        StreamPiece,
        StreamPiece,
    ]
    assert pieces[0] == ToolCallPiece(id="call_1", name="ls", arguments={})
    assert pieces[1] == ToolCallPiece(
        id="call_2", name="read", arguments={"path": "Homelab/aws-route53.md"}
    )
    assert pieces[3] == StreamPiece("content", "Done! ")
    # The read document is recorded for done.sources / query_log (task 04).
    assert holder.read_docs == [target]
    assert holder.tool_calls == 2

    # Phase 45: no per-tool budgets — the tools stay offered on every
    # request (the round cap, not spent budgets, bounds the loop), so
    # the answer request still carries them (2 rounds < default cap 10).
    assert llm.requests[0][1] == AGENT_TOOLS
    assert llm.requests[1][1] == AGENT_TOOLS
    assert llm.requests[2][1] == AGENT_TOOLS
    assert len(llm.requests) == 3

    # read splits the combined form at the FIRST slash — one exact
    # lookup, no self-correction candidates (phase 70).
    assert calls == [("Homelab", "aws-route53.md")]

    # The follow-up request carries the assistant tool-call + tool result.
    msgs = llm.requests[1][0]
    assert msgs[0] == {"role": "system", "content": "SYSTEM_PROMPT"}
    assert msgs[1] == {"role": "user", "content": "QUESTION"}
    assert msgs[2] == {
        "role": "assistant",
        "content": None,
        "tool_calls": [
            {
                "id": "call_1",
                "type": "function",
                "function": {"name": "ls", "arguments": "{}"},
            }
        ],
    }
    assert msgs[3] == {
        "role": "tool",
        "tool_call_id": "call_1",
        "content": (
            "2 documents:\n"
            "source: Deployments | path: backups.md | title: Backup Strategy\n"
            "source: Homelab | path: aws-route53.md | title: AWS Route53 Records"
        ),
    }
    # The second follow-up request carries the read call + the FULL text.
    msgs = llm.requests[2][0]
    assert msgs[4]["role"] == "assistant"
    assert msgs[4]["tool_calls"][0]["id"] == "call_2"
    assert json.loads(msgs[4]["tool_calls"][0]["function"]["arguments"]) == {
        "path": "Homelab/aws-route53.md"
    }
    assert msgs[5] == {
        "role": "tool",
        "tool_call_id": "call_2",
        "content": "Document Homelab/aws-route53.md:\nR53-CONTENT",  # full text, no cap
    }


def test_content_and_tool_call_in_one_stream_keeps_both(
    monkeypatch: pytest.MonkeyPatch,
) -> None:
    """Rare stream with content AND a tool call: the content stays (it was
    already emitted) and the tool still runs."""
    monkeypatch.setattr(agent, "list_catalog", lambda db: [])
    holder = AgentHolder()
    llm = ScriptedLLM(
        [
            StreamPiece("content", "Let me check "),
            ToolCallPiece(id="call_1", name="ls", arguments={}),
        ],
        [StreamPiece("content", "the answer")],
    )
    pieces = asyncio.run(_run(llm, holder, _settings()))
    assert [type(p) for p in pieces] == [StreamPiece, ToolCallPiece, StreamPiece]
    assert holder.tool_calls == 1  # the tool ran despite the content
    assert llm.requests[1][0][3]["content"] == "0 documents:\n"


# ---------- ls: full catalog + scoping ----------


def test_ls_full_catalog_format(monkeypatch: pytest.MonkeyPatch) -> None:
    """No argument: the full catalog in the phase-63 labeled-field format
    (``source: X | path: Y | title: Z``) — counted; no registry lookup."""
    catalog = [
        ("Deployments", "backups.md", "Backup Strategy"),
        ("Homelab", "aws-route53.md", "AWS Route53 Records"),
    ]
    monkeypatch.setattr(agent, "list_catalog", lambda db: catalog)

    def _boom_sources(*_a: Any, **_k: Any) -> None:
        raise AssertionError("no registry lookup for an unscoped ls")

    monkeypatch.setattr(agent, "list_source_names", _boom_sources)
    holder = AgentHolder()
    llm = ScriptedLLM(
        [ToolCallPiece(id="call_1", name="ls", arguments={})],
        [StreamPiece("content", "ans")],
    )
    asyncio.run(_run(llm, holder, _settings()))
    assert llm.requests[1][0][3]["content"] == (
        "2 documents:\n"
        "source: Deployments | path: backups.md | title: Backup Strategy\n"
        "source: Homelab | path: aws-route53.md | title: AWS Route53 Records"
    )
    assert holder.tool_calls == 1


def test_ls_empty_catalog_says_zero_documents(monkeypatch: pytest.MonkeyPatch) -> None:
    monkeypatch.setattr(agent, "list_catalog", lambda db: [])
    holder = AgentHolder()
    llm = ScriptedLLM(
        [ToolCallPiece(id="call_1", name="ls", arguments={})],
        [StreamPiece("content", "ans")],
    )
    asyncio.run(_run(llm, holder, _settings()))
    assert llm.requests[1][0][3]["content"] == "0 documents:\n"
    assert holder.tool_calls == 1


@pytest.mark.parametrize(
    ("arguments", "label"),
    [
        ({"path": "   "}, "blank path"),
        ({"path": 7}, "non-string path"),
    ],
)
def test_ls_blank_path_lists_full_catalog(
    monkeypatch: pytest.MonkeyPatch, arguments: dict[str, Any], label: str
) -> None:
    """A blank (or non-string) ``path`` is treated as omitted — the full
    catalog, counted (no refusal for an empty scope)."""
    catalog = [("S", "a.md", "A")]
    monkeypatch.setattr(agent, "list_catalog", lambda db: catalog)
    monkeypatch.setattr(agent, "list_source_names", lambda db: ["S"])
    holder = AgentHolder()
    llm = ScriptedLLM(
        [ToolCallPiece(id="call_1", name="ls", arguments=arguments)],
        [StreamPiece("content", "ans")],
    )
    asyncio.run(_run(llm, holder, _settings()))
    assert (
        llm.requests[1][0][3]["content"] == "1 documents:\nsource: S | path: a.md | title: A"
    )
    assert holder.tool_calls == 1


def test_ls_scoped_to_known_source(monkeypatch: pytest.MonkeyPatch) -> None:
    """A known source name: the same listing filtered to that source —
    counted."""
    catalog = [
        ("Deployments", "backups.md", "Backup Strategy"),
        ("Homelab", "a.md", "A"),
        ("Homelab", "b.md", "B"),
    ]
    monkeypatch.setattr(agent, "list_catalog", lambda db: catalog)
    monkeypatch.setattr(agent, "list_source_names", lambda db: ["Deployments", "Homelab"])
    holder = AgentHolder()
    llm = ScriptedLLM(
        [ToolCallPiece(id="call_1", name="ls", arguments={"path": "Homelab"})],
        [StreamPiece("content", "ans")],
    )
    asyncio.run(_run(llm, holder, _settings()))
    assert llm.requests[1][0][3]["content"] == (
        "2 documents:\n"
        "source: Homelab | path: a.md | title: A\n"
        "source: Homelab | path: b.md | title: B"
    )
    assert holder.tool_calls == 1


def test_ls_scoped_known_source_with_zero_docs_counts(
    monkeypatch: pytest.MonkeyPatch,
) -> None:
    """A registered source with no indexed documents is KNOWN (the
    registry is the source of truth, not the catalog): it lists as
    ``0 documents:`` — a valid, counted result, not a refusal."""
    monkeypatch.setattr(agent, "list_catalog", lambda db: [("Other", "a.md", "A")])
    monkeypatch.setattr(agent, "list_source_names", lambda db: ["Homelab", "Other"])
    holder = AgentHolder()
    llm = ScriptedLLM(
        [ToolCallPiece(id="call_1", name="ls", arguments={"path": "Homelab"})],
        [StreamPiece("content", "ans")],
    )
    asyncio.run(_run(llm, holder, _settings()))
    assert llm.requests[1][0][3]["content"] == "0 documents:\n"
    assert holder.tool_calls == 1  # an executed ls, not a refusal
    assert llm.requests[1][1] == AGENT_TOOLS


def test_ls_scoped_unknown_source_refused(monkeypatch: pytest.MonkeyPatch) -> None:
    """A ``path`` matching no source name is a refusal — not counted, the
    round cap bounds its repetition."""
    monkeypatch.setattr(agent, "list_catalog", lambda db: [("S", "a.md", "A")])
    monkeypatch.setattr(agent, "list_source_names", lambda db: ["S"])
    holder = AgentHolder()
    llm = ScriptedLLM(
        [ToolCallPiece(id="call_1", name="ls", arguments={"path": "Ghost"})],
        [StreamPiece("content", "ans")],
    )
    asyncio.run(_run(llm, holder, _settings()))
    assert holder.tool_calls == 0  # a refusal counts in nothing
    assert (
        llm.requests[1][0][3]["content"] == "No source named 'Ghost' — check the ls output."
    )
    assert llm.requests[1][1] == AGENT_TOOLS  # rejected → tools stay offered


# ---------- read: the canonical combined source/path form ----------


def test_read_combined_path_resolves_and_returns_full_content(
    monkeypatch: pytest.MonkeyPatch,
) -> None:
    """The combined ``source/path`` form (the model's trained shape) is
    the canonical identity: split at the FIRST '/', one exact lookup,
    the full content returned (A7-revised: never truncated) — even when
    the path itself carries further slashes."""
    doc = _doc("Homelab", "active/container_caddy/caddy.md", "Caddy", "CADDY-CONTENT")
    calls: list[tuple[str, str]] = []

    def _find(db: Any, source: str, path: str) -> Document | None:
        calls.append((source, path))
        return (
            doc
            if (source, path) == ("Homelab", "active/container_caddy/caddy.md")
            else None
        )

    monkeypatch.setattr(agent, "find_document", _find)
    holder = AgentHolder()
    llm = ScriptedLLM(
        [
            ToolCallPiece(
                id="call_1",
                name="read",
                arguments={"path": "Homelab/active/container_caddy/caddy.md"},
            )
        ],
        [StreamPiece("content", "ans")],
    )
    asyncio.run(_run(llm, holder, _settings()))
    # First-slash split — exactly one lookup, the canonical pair.
    assert calls == [("Homelab", "active/container_caddy/caddy.md")]
    assert holder.read_docs == [doc]
    assert holder.tool_calls == 1
    assert llm.requests[1][0][3]["content"] == (
        "Document Homelab/active/container_caddy/caddy.md:\nCADDY-CONTENT"
    )


def test_read_bare_source_name_refused_without_db(monkeypatch: pytest.MonkeyPatch) -> None:
    """A bare source name (no '/') can never be a document — the
    no-document refusal (the argument echoed as passed), no DB lookup,
    nothing counted."""
    monkeypatch.setattr(agent, "list_catalog", lambda db: [("Homelab", "a.md", "A")])

    def _boom(*_a: Any, **_k: Any) -> None:
        raise AssertionError("find_document must not run for a bare source name")

    monkeypatch.setattr(agent, "find_document", _boom)
    holder = AgentHolder()
    llm = ScriptedLLM(
        [ToolCallPiece(id="call_1", name="read", arguments={"path": "Homelab"})],
        [StreamPiece("content", "ans")],
    )
    asyncio.run(_run(llm, holder, _settings()))
    assert holder.read_docs == [] and holder.tool_calls == 0
    assert llm.requests[1][0][3]["content"] == (
        "No document at 'Homelab' — check the ls output."
    )
    assert llm.requests[1][1] == AGENT_TOOLS


def test_read_unknown_path_refused_echoing_argument(
    monkeypatch: pytest.MonkeyPatch,
) -> None:
    """An unknown combined identity → the refusal echoing the argument as
    passed (the model sees its own form) — the old split-teaching refusal
    is gone (phase 70)."""
    monkeypatch.setattr(agent, "find_document", lambda db, source, path: None)
    holder = AgentHolder()
    llm = ScriptedLLM(
        [ToolCallPiece(id="call_1", name="read", arguments={"path": "S/ghost.md"})],
        [StreamPiece("content", "ans")],
    )
    asyncio.run(_run(llm, holder, _settings()))
    assert holder.read_docs == [] and holder.tool_calls == 0
    assert llm.requests[1][0][3]["content"] == (
        "No document at 'S/ghost.md' — check the ls output."
    )
    assert llm.requests[1][1] == AGENT_TOOLS  # tools stay offered (cap bounds)


@pytest.mark.parametrize(
    ("arguments", "label"),
    [
        ({}, "no arguments"),
        ({"path": ""}, "empty path"),
        ({"path": "   "}, "blank path"),
        ({"path": 7}, "non-string path"),
        ({"path": None}, "null path"),
    ],
)
def test_read_missing_arguments_refused(
    monkeypatch: pytest.MonkeyPatch, arguments: dict[str, Any], label: str
) -> None:
    def _boom(*_a: Any, **_k: Any) -> None:
        raise AssertionError(f"find_document must not be called ({label})")

    monkeypatch.setattr(agent, "find_document", _boom)
    holder = AgentHolder()
    llm = ScriptedLLM(
        [ToolCallPiece(id="call_1", name="read", arguments=arguments)],
        [StreamPiece("content", "ans")],
    )
    asyncio.run(_run(llm, holder, _settings()))
    assert holder.read_docs == [] and holder.tool_calls == 0
    assert llm.requests[1][0][3]["content"] == agent.MISSING_READ_ARGS
    assert llm.requests[1][1] == AGENT_TOOLS


def test_reading_a_seed_doc_is_already_in_context(monkeypatch: pytest.MonkeyPatch) -> None:
    """The combined identity of a seeded document: its split pair is in
    the known set → ALREADY_IN_CONTEXT with no DB lookup (the dedupe
    check precedes the resolve)."""
    seed = [_doc("Homelab", "kubernetes.md", "Kubernetes", "K8S-CONTENT")]

    def _boom(*_a: Any, **_k: Any) -> None:
        raise AssertionError("find_document must not be called for a seeded doc")

    monkeypatch.setattr(agent, "find_document", _boom)
    holder = AgentHolder()
    llm = ScriptedLLM(
        [
            ToolCallPiece(
                id="call_1",
                name="read",
                arguments={"path": "Homelab/kubernetes.md"},
            )
        ],
        [StreamPiece("content", "ans")],
    )
    asyncio.run(_run(llm, holder, _settings(), seed_docs=seed))
    assert holder.read_docs == [] and holder.tool_calls == 0
    assert llm.requests[1][0][3]["content"] == agent.ALREADY_IN_CONTEXT
    # Rejected → the tools are still offered on the next request (the
    # round cap is the only bound).
    assert llm.requests[1][1] == AGENT_TOOLS


def test_reading_an_already_read_doc_is_deduped(monkeypatch: pytest.MonkeyPatch) -> None:
    """The second read of the same document (holder.read_docs) →
    ALREADY_IN_CONTEXT — appended once, counted once."""
    doc = _doc("S", "a.md", "A", "A-CONTENT")
    monkeypatch.setattr(agent, "find_document", lambda db, source, path: doc)
    holder = AgentHolder()
    llm = ScriptedLLM(
        [ToolCallPiece(id="call_1", name="read", arguments={"path": "S/a.md"})],
        [ToolCallPiece(id="call_2", name="read", arguments={"path": "S/a.md"})],
        [StreamPiece("content", "ans")],
    )
    asyncio.run(_run(llm, holder, _settings()))
    assert holder.read_docs == [doc]  # appended exactly once
    assert holder.tool_calls == 1  # the re-read counts nothing
    assert llm.requests[1][0][3]["content"] == "Document S/a.md:\nA-CONTENT"
    assert llm.requests[2][0][5]["content"] == agent.ALREADY_IN_CONTEXT
    # Rejected → the tools are still offered on the next request…
    assert llm.requests[2][1] == AGENT_TOOLS


def test_unknown_tool_name_refused(monkeypatch: pytest.MonkeyPatch) -> None:
    monkeypatch.setattr(agent, "list_catalog", lambda db: [])
    holder = AgentHolder()
    llm = ScriptedLLM(
        [ToolCallPiece(id="call_1", name="delete_universe", arguments={"x": 1})],
        [StreamPiece("content", "ans")],
    )
    asyncio.run(_run(llm, holder, _settings()))
    assert holder.read_docs == [] and holder.tool_calls == 0
    assert llm.requests[1][0][3]["content"] == agent.UNKNOWN_TOOL
    assert llm.requests[1][1] == AGENT_TOOLS  # rejected → tools stay offered


# ---------- grep (the phase-68 A5 contract under the new name) ----------


def test_grep_document_case_insensitive_line_numbers() -> None:
    """Case-insensitive fixed substring, 1-based line numbers, file order,
    repeated matches within a line collapse to one match (grep semantics)."""
    content = "The NEEDLE is here\nno hit\nneedle again\nNEEDLE NEEDLE\n"
    assert agent.grep_document(content, "NEEDLE") == [
        (1, "The NEEDLE is here"),
        (3, "needle again"),
        (4, "NEEDLE NEEDLE"),
    ]


def test_grep_document_rstrips_lines_and_empty_content() -> None:
    assert agent.grep_document("hello \t\nworld ", "WORLD") == [(2, "world")]
    assert agent.grep_document("", "x") == []
    assert agent.grep_document("no newlines", "NO") == [(1, "no newlines")]
    assert agent.grep_document("a\nb\n", "MISSING") == []


def test_grep_whole_kb_grep_style_output(monkeypatch: pytest.MonkeyPatch) -> None:
    """Whole-KB grep: catalog order, `source/path:line: text` lines,
    case-insensitive; the call counts in ``tool_calls`` and never touches
    ``read_docs``; the tools stay offered on the answer request."""
    d1 = _doc("Alpha", "a/one.md", "One", "first\nNEEDLE in one\nlast")
    d2 = _doc("Beta", "b/two.md", "Two", "no hit\nneedle in two\n")
    monkeypatch.setattr(agent, "all_documents", lambda db: [d1, d2])
    holder = AgentHolder()
    llm = ScriptedLLM(
        [ToolCallPiece(id="call_1", name="grep", arguments={"pattern": "needle"})],
        [StreamPiece("content", "ans")],
    )
    asyncio.run(_run(llm, holder, _settings()))
    assert llm.requests[1][0][3]["content"] == (
        "Alpha/a/one.md:2: NEEDLE in one\n"
        "Beta/b/two.md:2: needle in two"
    )
    assert holder.tool_calls == 1
    assert holder.read_docs == []  # locked A5: a grep adds no context
    assert llm.requests[1][1] == AGENT_TOOLS  # tools stay offered


def test_grep_capped_at_20_matches_in_catalog_order(
    monkeypatch: pytest.MonkeyPatch,
) -> None:
    """The 20-match cap is GLOBAL across documents in catalog order, and
    the scan stops once it is hit (a 35-match corpus yields exactly 20)."""
    d1 = _doc("S", "a.md", "A", "\n".join(f"hit-{i}" for i in range(15)))
    d2 = _doc("S", "b.md", "B", "\n".join(f"hit-{i}" for i in range(20)))
    monkeypatch.setattr(agent, "all_documents", lambda db: [d1, d2])
    holder = AgentHolder()
    llm = ScriptedLLM(
        [ToolCallPiece(id="call_1", name="grep", arguments={"pattern": "hit-"})],
        [StreamPiece("content", "ans")],
    )
    asyncio.run(_run(llm, holder, _settings()))
    lines = llm.requests[1][0][3]["content"].split("\n")
    assert len(lines) == agent.SEARCH_MAX_MATCHES
    assert lines[0] == "S/a.md:1: hit-0"
    assert lines[14] == "S/a.md:15: hit-14"  # all of a.md
    assert lines[15] == "S/b.md:1: hit-0"  # then b.md, in order
    assert lines[19] == "S/b.md:5: hit-4"  # cut at the global cap
    assert holder.tool_calls == 1


def test_grep_truncates_match_lines_at_200_chars(monkeypatch: pytest.MonkeyPatch) -> None:
    """A 300-char match line yields exactly 200 chars of it (no crash)."""
    d1 = _doc("S", "a.md", "A", "top\n" + "x" * 300 + " NEEDLE tail")
    monkeypatch.setattr(agent, "all_documents", lambda db: [d1])
    holder = AgentHolder()
    llm = ScriptedLLM(
        [ToolCallPiece(id="call_1", name="grep", arguments={"pattern": "needle"})],
        [StreamPiece("content", "ans")],
    )
    asyncio.run(_run(llm, holder, _settings()))
    assert llm.requests[1][0][3]["content"] == f"S/a.md:2: {'x' * agent.SEARCH_LINE_LIMIT}"
    assert holder.tool_calls == 1


def test_grep_scoped_to_one_document(monkeypatch: pytest.MonkeyPatch) -> None:
    """Scoped grep: only the named document is loaded (find_document on
    the first-slash split), ``all_documents`` never runs, and the match
    line carries its path."""
    d1 = _doc("S", "a.md", "A", "needle here")

    def _find(db: Any, source: str, path: str) -> Document | None:
        if (source, path) == ("S", "a.md"):
            return d1
        raise AssertionError(
            f"find_document({source}, {path}) — the scoped "
            "grep must not load any other document"
        )

    def _boom(*_a: Any, **_k: Any) -> None:
        raise AssertionError("all_documents must not run for a scoped grep")

    monkeypatch.setattr(agent, "find_document", _find)
    monkeypatch.setattr(agent, "all_documents", _boom)
    holder = AgentHolder()
    llm = ScriptedLLM(
        [
            ToolCallPiece(
                id="call_1",
                name="grep",
                arguments={"pattern": "needle", "path": "S/a.md"},
            )
        ],
        [StreamPiece("content", "ans")],
    )
    asyncio.run(_run(llm, holder, _settings()))
    assert llm.requests[1][0][3]["content"] == "S/a.md:1: needle here"
    assert holder.tool_calls == 1
    assert holder.read_docs == []  # grepped doc did not enter the context


def test_grep_scoped_combined_path_with_nested_path(
    monkeypatch: pytest.MonkeyPatch,
) -> None:
    """A combined target whose path itself contains slashes: the split is
    at the FIRST slash — the scoped grep runs on the right document."""
    d1 = _doc("S", "deep/nested/a.md", "A", "needle here")

    def _find(db: Any, source: str, path: str) -> Document | None:
        if (source, path) == ("S", "deep/nested/a.md"):
            return d1
        raise AssertionError(f"find_document({source}, {path}) — wrong first-slash split")

    monkeypatch.setattr(agent, "find_document", _find)
    holder = AgentHolder()
    llm = ScriptedLLM(
        [
            ToolCallPiece(
                id="call_1",
                name="grep",
                arguments={"pattern": "needle", "path": "S/deep/nested/a.md"},
            )
        ],
        [StreamPiece("content", "ans")],
    )
    asyncio.run(_run(llm, holder, _settings()))
    assert llm.requests[1][0][3]["content"] == "S/deep/nested/a.md:1: needle here"
    assert holder.tool_calls == 1


def test_grep_scoped_missing_document_refused(monkeypatch: pytest.MonkeyPatch) -> None:
    monkeypatch.setattr(agent, "find_document", lambda db, source, path: None)
    holder = AgentHolder()
    llm = ScriptedLLM(
        [
            ToolCallPiece(
                id="call_1",
                name="grep",
                arguments={"pattern": "x", "path": "S/ghost.md"},
            )
        ],
        [StreamPiece("content", "ans")],
    )
    asyncio.run(_run(llm, holder, _settings()))
    assert llm.requests[1][0][3]["content"] == (
        "No document at 'S/ghost.md' — check the ls output."
    )
    assert holder.tool_calls == 0 and holder.read_docs == []  # a refusal
    assert llm.requests[1][1] == AGENT_TOOLS  # rejected → tools stay offered


def test_grep_scoped_bare_source_name_refused_without_db(
    monkeypatch: pytest.MonkeyPatch,
) -> None:
    """A bare source name as the grep target can never resolve to exactly
    one document — the no-document refusal, no DB lookup, not counted."""
    def _boom(*_a: Any, **_k: Any) -> None:
        raise AssertionError("find_document must not run for a bare source name")

    monkeypatch.setattr(agent, "find_document", _boom)
    holder = AgentHolder()
    llm = ScriptedLLM(
        [
            ToolCallPiece(
                id="call_1",
                name="grep",
                arguments={"pattern": "x", "path": "Homelab"},
            )
        ],
        [StreamPiece("content", "ans")],
    )
    asyncio.run(_run(llm, holder, _settings()))
    assert llm.requests[1][0][3]["content"] == (
        "No document at 'Homelab' — check the ls output."
    )
    assert holder.tool_calls == 0 and holder.read_docs == []


@pytest.mark.parametrize(
    ("arguments", "label"),
    [
        ({}, "no arguments"),
        ({"pattern": ""}, "empty pattern"),
        ({"pattern": "   "}, "whitespace pattern"),
        ({"pattern": 42}, "non-string pattern"),
        ({"pattern": None}, "null pattern"),
    ],
)
def test_grep_missing_arguments_refused(
    monkeypatch: pytest.MonkeyPatch, arguments: dict[str, Any], label: str
) -> None:
    """A missing/blank/non-string pattern → the missing-args refusal, with
    no DB access at all."""

    def _boom(*_a: Any, **_k: Any) -> None:
        raise AssertionError(f"no DB access for a refused grep ({label})")

    monkeypatch.setattr(agent, "all_documents", _boom)
    monkeypatch.setattr(agent, "find_document", _boom)
    holder = AgentHolder()
    llm = ScriptedLLM(
        [ToolCallPiece(id="call_1", name="grep", arguments=arguments)],
        [StreamPiece("content", "ans")],
    )
    asyncio.run(_run(llm, holder, _settings()))
    assert llm.requests[1][0][3]["content"] == agent.MISSING_SEARCH_ARGS
    assert holder.tool_calls == 0 and holder.read_docs == []
    assert llm.requests[1][1] == AGENT_TOOLS  # rejected → tools stay offered


def test_grep_no_matches_whole_kb(monkeypatch: pytest.MonkeyPatch) -> None:
    """Zero hits across the KB → the no-match line (pattern quoted); the
    grep still executed, so it counts — and never adds context."""
    monkeypatch.setattr(
        agent, "all_documents", lambda db: [_doc("S", "a.md", "A", "nothing here")]
    )
    holder = AgentHolder()
    llm = ScriptedLLM(
        [ToolCallPiece(id="call_1", name="grep", arguments={"pattern": "zebra"})],
        [StreamPiece("content", "ans")],
    )
    asyncio.run(_run(llm, holder, _settings()))
    assert llm.requests[1][0][3]["content"] == (
        "No matches for 'zebra' in the knowledge base."
    )
    assert holder.tool_calls == 1
    assert holder.read_docs == []


def test_grep_no_matches_scoped(monkeypatch: pytest.MonkeyPatch) -> None:
    """The scoped no-match line is keyed on the resolved source/path."""
    doc = _doc("S", "a.md", "A", "nothing here")

    def _find(db: Any, source: str, path: str) -> Document | None:
        return doc if (source, path) == ("S", "a.md") else None

    monkeypatch.setattr(agent, "find_document", _find)
    holder = AgentHolder()
    llm = ScriptedLLM(
        [
            ToolCallPiece(
                id="call_1",
                name="grep",
                arguments={"pattern": "zebra", "path": "S/a.md"},
            )
        ],
        [StreamPiece("content", "ans")],
    )
    asyncio.run(_run(llm, holder, _settings()))
    assert llm.requests[1][0][3]["content"] == "No matches for 'zebra' in S/a.md."
    assert holder.tool_calls == 1
    assert holder.read_docs == []


def test_grep_no_match_truncates_long_pattern(monkeypatch: pytest.MonkeyPatch) -> None:
    """A pattern longer than 100 chars is truncated in the no-match line
    (kept short); the grep itself still runs on the full pattern."""
    monkeypatch.setattr(agent, "all_documents", lambda db: [])
    holder = AgentHolder()
    llm = ScriptedLLM(
        [
            ToolCallPiece(id="call_1", name="grep", arguments={"pattern": "p" * 150})
        ],
        [StreamPiece("content", "ans")],
    )
    asyncio.run(_run(llm, holder, _settings()))
    assert llm.requests[1][0][3]["content"] == (
        f"No matches for '{'p' * 100}' in the knowledge base."
    )
    assert holder.tool_calls == 1


def test_grep_counts_but_never_adds_context(monkeypatch: pytest.MonkeyPatch) -> None:
    """The locate-then-read workflow: a grep finds the document but does
    NOT add it — the subsequent read does (and is not rejected as
    already-in-context, because the grep touched nothing)."""
    doc = _doc("S", "a.md", "A", "needle here")
    monkeypatch.setattr(agent, "all_documents", lambda db: [doc])
    monkeypatch.setattr(agent, "find_document", lambda db, source, path: doc)
    holder = AgentHolder()
    llm = ScriptedLLM(
        [ToolCallPiece(id="call_1", name="grep", arguments={"pattern": "needle"})],
        [ToolCallPiece(id="call_2", name="read", arguments={"path": "S/a.md"})],
        [StreamPiece("content", "ans")],
    )
    asyncio.run(_run(llm, holder, _settings()))
    assert holder.tool_calls == 2  # grep + read, both executed
    assert holder.read_docs == [doc]  # only the read added context (A5)
    assert llm.requests[1][0][3]["content"] == "S/a.md:1: needle here"
    assert llm.requests[2][0][5]["content"] == "Document S/a.md:\nneedle here"


# ---------- unlimited calls: re-lists and multi-reads (phase 45) ----------


def test_relist_executes_and_counts(monkeypatch: pytest.MonkeyPatch) -> None:
    """Re-lists execute — a second ``ls`` in one turn returns the catalog
    again and counts in ``tool_calls`` (no budget to exhaust)."""
    catalog = [
        ("Deployments", "backups.md", "Backup Strategy"),
        ("Homelab", "aws-route53.md", "AWS Route53 Records"),
    ]
    monkeypatch.setattr(agent, "list_catalog", lambda db: catalog)
    holder = AgentHolder()
    llm = ScriptedLLM(
        [ToolCallPiece(id="call_1", name="ls", arguments={})],
        [ToolCallPiece(id="call_2", name="ls", arguments={})],
        [StreamPiece("content", "ans")],
    )
    asyncio.run(_run(llm, holder, _settings()))
    assert holder.tool_calls == 2  # both re-lists executed and counted
    listing = (
        "2 documents:\n"
        "source: Deployments | path: backups.md | title: Backup Strategy\n"
        "source: Homelab | path: aws-route53.md | title: AWS Route53 Records"
    )
    # The answer request carries the catalog a second time as a tool result.
    assert llm.requests[2][0][3]["content"] == listing  # first listing
    assert llm.requests[2][0][5]["content"] == listing  # the re-list
    assert llm.requests[2][1] == AGENT_TOOLS  # still offered (no budgets)


def test_multi_read_executes_without_budgets(monkeypatch: pytest.MonkeyPatch) -> None:
    """Reads are no longer budgeted either — two different documents can
    be read in one turn (re-reading the same one is still deduped via
    ALREADY_IN_CONTEXT — see the rejection tests)."""
    a = _doc("S", "a.md", "A", "A-CONTENT")
    b = _doc("S", "b.md", "B", "B-CONTENT")
    monkeypatch.setattr(
        agent, "find_document", lambda db, source, path: {"a.md": a, "b.md": b}[path]
    )
    holder = AgentHolder()
    llm = ScriptedLLM(
        [ToolCallPiece(id="call_1", name="read", arguments={"path": "S/a.md"})],
        [ToolCallPiece(id="call_2", name="read", arguments={"path": "S/b.md"})],
        [StreamPiece("content", "ans")],
    )
    asyncio.run(_run(llm, holder, _settings()))
    assert holder.read_docs == [a, b]  # both reads appended, in order
    assert holder.tool_calls == 2
    assert llm.requests[1][0][3]["content"] == "Document S/a.md:\nA-CONTENT"
    assert llm.requests[2][0][5]["content"] == "Document S/b.md:\nB-CONTENT"
    assert llm.requests[2][1] == AGENT_TOOLS  # the second read was still offered


# ---------- round cap (phase 45: replaces the per-tool budgets) ----------


def test_always_ls_bounded_by_round_cap(monkeypatch: pytest.MonkeyPatch) -> None:
    """A model that keeps calling ``ls`` gets exactly
    ``agent_max_rounds`` tool rounds, then one forced ``tools=None``
    request streams the answer — the cap is the only forced exit."""
    monkeypatch.setattr(agent, "list_catalog", lambda db: [("S", "a.md", "A")])
    listing = "1 documents:\nsource: S | path: a.md | title: A"
    holder = AgentHolder()
    llm = ScriptedLLM(
        [ToolCallPiece(id="call_1", name="ls", arguments={})],
        [ToolCallPiece(id="call_2", name="ls", arguments={})],
        [ToolCallPiece(id="call_3", name="ls", arguments={})],
        [StreamPiece("content", "forced answer")],
    )
    pieces = asyncio.run(_run(llm, holder, _settings(agent_max_rounds=3)))
    assert [type(p) for p in pieces] == [
        ToolCallPiece,
        ToolCallPiece,
        ToolCallPiece,
        StreamPiece,
    ]
    assert len(llm.requests) == 4  # 3 tool rounds + the forced answer
    # The three tool rounds were offered the tools…
    assert llm.requests[0][1] == AGENT_TOOLS
    assert llm.requests[1][1] == AGENT_TOOLS
    assert llm.requests[2][1] == AGENT_TOOLS
    # …and the forced final request carries no tools, whatever is left.
    assert llm.requests[3][1] is None
    # Every re-list executed and counted.
    assert holder.tool_calls == 3
    # The final request carries all three executed listings as history.
    final_msgs = llm.requests[3][0]
    assert len(final_msgs) == 8  # 2 + 3 rounds × (assistant + tool)
    assert final_msgs[3]["content"] == listing
    assert final_msgs[5]["content"] == listing
    assert final_msgs[7]["content"] == listing


def test_zero_max_rounds_is_one_request_without_tools() -> None:
    """``agent_max_rounds=0`` — the kill switch: exactly one request,
    ``tools=None``, no tool lines, no history growth (byte-identical to
    the pre-phase-37 path)."""
    holder = AgentHolder()
    llm = ScriptedLLM([StreamPiece("thinking", "t "), StreamPiece("content", "direct answer")])
    pieces = asyncio.run(_run(llm, holder, _settings(agent_max_rounds=0)))
    assert [type(p) for p in pieces] == [StreamPiece, StreamPiece]
    assert len(llm.requests) == 1
    assert llm.requests[0][1] is None
    assert llm.requests[0][0] == [
        {"role": "system", "content": "SYSTEM_PROMPT"},
        {"role": "user", "content": "QUESTION"},
    ]
    assert holder.read_docs == [] and holder.tool_calls == 0


def test_rejected_read_spam_runs_to_round_cap(monkeypatch: pytest.MonkeyPatch) -> None:
    """Every call rejected (unknown document — "No document at …"):
    rejections no longer end the loop early via budgets — the round cap
    bounds them and forces the final no-tools answer."""
    monkeypatch.setattr(agent, "find_document", lambda db, source, path: None)
    holder = AgentHolder()
    llm = ScriptedLLM(
        [ToolCallPiece(id="call_1", name="read", arguments={"path": "S/ghost.md"})],
        [ToolCallPiece(id="call_2", name="read", arguments={"path": "S/ghost.md"})],
        [ToolCallPiece(id="call_3", name="read", arguments={"path": "S/ghost.md"})],
        [StreamPiece("content", "forced answer")],
    )
    asyncio.run(_run(llm, holder, _settings(agent_max_rounds=3)))
    assert len(llm.requests) == 4  # 3 rejected rounds + the forced answer
    assert llm.requests[0][1] == AGENT_TOOLS
    assert llm.requests[1][1] == AGENT_TOOLS
    assert llm.requests[2][1] == AGENT_TOOLS
    assert llm.requests[3][1] is None  # the forced final request: no tools
    assert holder.read_docs == [] and holder.tool_calls == 0  # nothing executed
    refusal = "No document at 'S/ghost.md' — check the ls output."
    assert llm.requests[1][0][3]["content"] == refusal
    assert llm.requests[2][0][5]["content"] == refusal
    assert llm.requests[3][0][7]["content"] == refusal


# ---------- retries inside the agent loop (phase 67, locked A2) ----------


class FailingLLM:
    """A scripted fake whose Nth ``chat_stream`` call yields pieces and
    then raises (phase 67): ``attempts`` is a list of ``(pieces, error)``
    — an error after zero pieces = "the endpoint died before the first
    token"; after some pieces = a mid-stream drop. Records every
    request's messages/tools and the indices of the attempts whose stream
    teardown ran (``closed``)."""

    def __init__(
        self,
        attempts: list[tuple[list[StreamPiece | ToolCallPiece], Exception | None]],
    ) -> None:
        self.attempts = list(attempts)
        self.requests: list[tuple[list[dict[str, Any]], list[dict[str, Any]] | None]] = []
        #: Indices of attempts whose stream teardown has run.
        self.closed: list[int] = []

    def chat_stream(
        self,
        messages: list[dict[str, str]],
        tools: list[dict[str, Any]] | None = None,
        scaffolding: ScaffoldingFilter | None = None,  # phase 71: fed like the real client
    ) -> AsyncIterator[StreamPiece | ToolCallPiece]:
        index = len(self.requests)
        pieces, error = (
            self.attempts[index]
            if index < len(self.attempts)
            else ([], LLMError("script exhausted"))
        )
        self.requests.append(
            (deepcopy(messages), deepcopy(tools) if tools is not None else None)
        )
        return self._attempt(index, pieces, error, scaffolding)

    async def _attempt(
        self,
        index: int,
        pieces: list[StreamPiece | ToolCallPiece],
        error: Exception | None,
        scaffolding: ScaffoldingFilter | None,
    ) -> AsyncIterator[StreamPiece | ToolCallPiece]:
        try:
            for piece in pieces:
                if (
                    scaffolding is not None
                    and isinstance(piece, StreamPiece)
                    and piece.kind == "content"
                ):
                    cleaned = scaffolding.feed(piece.text)
                    if cleaned:
                        yield StreamPiece("content", cleaned)
                else:
                    yield piece
            if error is not None:
                # The tail is NOT flushed on a failed attempt — the real
                # client only flushes a cleanly completed stream.
                raise error
            if scaffolding is not None:
                tail = scaffolding.flush()
                if tail:
                    yield StreamPiece("content", tail)
        finally:
            self.closed.append(index)


def _record_sleeps(monkeypatch: pytest.MonkeyPatch) -> list[float]:
    """Monkeypatch ``asyncio.sleep`` (what ``chat_stream_retried`` awaits
    for the flat pre-retry delay) and record every awaited delay."""
    sleeps: list[float] = []

    async def fake_sleep(seconds: float) -> None:
        sleeps.append(seconds)

    monkeypatch.setattr(asyncio, "sleep", fake_sleep)
    return sleeps


def test_round_retried_before_first_piece(
    monkeypatch: pytest.MonkeyPatch,
    caplog: pytest.LogCaptureFixture,
) -> None:
    """A tool round that dies before its first piece is restarted with the
    same messages: the stream carries a RetryPiece BEFORE the tool call,
    the tool executes, the final answer streams, and the per-call log line
    is still emitted exactly once (retries are invisible to the loop)."""
    monkeypatch.setattr(agent, "list_catalog", lambda db: [("S", "a.md", "A")])
    holder = AgentHolder()
    llm = FailingLLM(
        [
            ([], LLMError("connection refused")),
            ([ToolCallPiece(id="call_1", name="ls", arguments={})], None),
            ([StreamPiece("content", "Done!")], None),
        ]
    )
    sleeps = _record_sleeps(monkeypatch)
    with caplog.at_level(logging.INFO, logger="app.agent"):
        pieces = asyncio.run(
            _run(llm, holder, _settings(agent_max_rounds=2, llm_retry_delay=2.5))
        )
    assert pieces == [
        RetryPiece(2, 4),  # default llm_retries=3 → 4 attempts
        ToolCallPiece(id="call_1", name="ls", arguments={}),
        StreamPiece("content", "Done!"),
    ]
    assert holder.tool_calls == 1
    assert holder.read_docs == []
    # The restart is byte-identical: same messages, same tools offered.
    assert len(llm.requests) == 3
    assert llm.requests[0] == llm.requests[1]
    assert llm.requests[0][1] == AGENT_TOOLS
    assert llm.requests[2][1] == AGENT_TOOLS  # the answer round still offered
    # The flat delay was awaited exactly once, before the retry.
    assert sleeps == [2.5]
    tool_logs = [r for r in caplog.records if r.getMessage().startswith("agent tool=")]
    assert len(tool_logs) == 1  # the retry did not re-run the tool or log
    assert tool_logs[0].getMessage() == "agent tool=ls args={} round=1/2"


def test_round_failure_after_first_piece_is_terminal(monkeypatch: pytest.MonkeyPatch) -> None:
    """Locked A2: a round that already streamed a piece fails the turn —
    the LLMError propagates out of ``run_agent``, no RetryPiece, no
    sleep, no second request, and the holder is untouched (the tool
    never ran)."""
    monkeypatch.setattr(agent, "list_catalog", lambda db: [])
    holder = AgentHolder()
    llm = FailingLLM(
        [([StreamPiece("content", "partial ")], LLMError("mid-stream drop"))]
    )
    sleeps = _record_sleeps(monkeypatch)

    async def drain() -> list[StreamPiece | ToolCallPiece | RetryPiece]:
        out: list[StreamPiece | ToolCallPiece | RetryPiece] = []
        with pytest.raises(LLMError, match="mid-stream drop"):
            async for piece in run_agent(
                cast("LLMClient", llm),
                cast("Session", None),
                system_prompt="SYSTEM_PROMPT",
                user_message="QUESTION",
                seed_docs=[],
                settings=_settings(),
                holder=holder,
            ):
                out.append(piece)
        return out

    out = asyncio.run(drain())
    assert out == [StreamPiece("content", "partial ")]  # no RetryPiece
    assert len(llm.requests) == 1  # no retry
    assert sleeps == []
    assert holder.read_docs == [] and holder.tool_calls == 0


def test_forced_final_no_tools_call_is_retried(monkeypatch: pytest.MonkeyPatch) -> None:
    """The forced final request (round cap reached) goes through the same
    retry rule: a failure before its first piece yields a RetryPiece and
    restarts with ``tools=None``; the answer from the retry streams."""
    monkeypatch.setattr(agent, "list_catalog", lambda db: [("S", "a.md", "A")])
    holder = AgentHolder()
    llm = FailingLLM(
        [
            ([ToolCallPiece(id="call_1", name="ls", arguments={})], None),
            ([ToolCallPiece(id="call_2", name="ls", arguments={})], None),
            ([], LLMError("down at the cap")),
            ([StreamPiece("content", "forced answer")], None),
        ]
    )
    pieces = asyncio.run(_run(llm, holder, _settings(agent_max_rounds=2)))
    assert [type(p) for p in pieces] == [
        ToolCallPiece,
        ToolCallPiece,
        RetryPiece,
        StreamPiece,
    ]
    assert pieces[2] == RetryPiece(2, 4)
    assert pieces[3] == StreamPiece("content", "forced answer")
    assert len(llm.requests) == 4  # 2 tool rounds + the final + its retry
    # The forced final (and its retry) carry no tools, whatever is left.
    assert llm.requests[2][1] is None
    assert llm.requests[3][1] is None
    # …and the restart is byte-identical.
    assert llm.requests[2][0] == llm.requests[3][0]
    assert holder.tool_calls == 2


def test_zero_retries_is_one_plain_attempt(monkeypatch: pytest.MonkeyPatch) -> None:
    """The kill-switch path (``llm_retries=0``): a dead round raises
    immediately — one request, no RetryPiece, no sleep (pre-phase-67
    behavior)."""
    monkeypatch.setattr(agent, "list_catalog", lambda db: [])
    holder = AgentHolder()
    llm = FailingLLM([([], LLMError("connection refused"))])
    sleeps = _record_sleeps(monkeypatch)

    async def drain() -> list[StreamPiece | ToolCallPiece | RetryPiece]:
        out: list[StreamPiece | ToolCallPiece | RetryPiece] = []
        with pytest.raises(LLMError, match="connection refused"):
            async for piece in run_agent(
                cast("LLMClient", llm),
                cast("Session", None),
                system_prompt="SYSTEM_PROMPT",
                user_message="QUESTION",
                seed_docs=[],
                settings=_settings(llm_retries=0),
                holder=holder,
            ):
                out.append(piece)
        return out

    out = asyncio.run(drain())
    assert out == []  # nothing streamed, no RetryPiece
    assert len(llm.requests) == 1
    assert sleeps == []
    assert holder.read_docs == [] and holder.tool_calls == 0


def test_abandon_mid_retry_sleep_leaks_nothing(monkeypatch: pytest.MonkeyPatch) -> None:
    """Consumer abandon while a retried round is parked in the pre-retry
    sleep (client disconnect): the driving task is cancelled cleanly, the
    production teardown ``aclose()`` on ``run_agent`` does not raise, the
    inner attempt's stream was torn down, and the retry never starts."""
    entered = asyncio.Event()

    async def parking_sleep(seconds: float) -> None:
        entered.set()
        await asyncio.Event().wait()  # park until the abandon arrives

    monkeypatch.setattr(asyncio, "sleep", parking_sleep)
    monkeypatch.setattr(agent, "list_catalog", lambda db: [])
    holder = AgentHolder()
    llm = FailingLLM(
        [([], LLMError("endpoint down")), ([StreamPiece("content", "never")], None)]
    )

    async def run() -> None:
        gen = run_agent(
            cast("LLMClient", llm),
            cast("Session", None),
            system_prompt="SYSTEM_PROMPT",
            user_message="QUESTION",
            seed_docs=[],
            settings=_settings(),
            holder=holder,
        )

        async def consumer() -> list[StreamPiece | ToolCallPiece | RetryPiece]:
            return [p async for p in gen]

        task = asyncio.ensure_future(consumer())
        await entered.wait()  # the round's retry is parked in the sleep
        assert not task.done()
        task.cancel()  # client disconnect: the driving task is cancelled
        with pytest.raises(asyncio.CancelledError):
            await task
        # Production teardown (phase 48 pattern): must not raise. ``run_agent``
        # is an async generator despite its AsyncIterator annotation.
        await cast(
            "AsyncGenerator[StreamPiece | ToolCallPiece | RetryPiece, None]", gen
        ).aclose()

    asyncio.run(run())
    assert len(llm.requests) == 1  # the retry never started
    assert llm.closed == [0]  # attempt 1's inner stream was torn down
    assert holder.read_docs == [] and holder.tool_calls == 0


def test_retries_are_invisible_to_the_round_cap(
    monkeypatch: pytest.MonkeyPatch,
    caplog: pytest.LogCaptureFixture,
) -> None:
    """A failing-then-succeeding round consumes ONE round: with a cap of
    2, the retried first round and the second tool round fill the cap —
    the forced final follows the SECOND call, and the log lines read
    round=1/2 and round=2/2."""
    monkeypatch.setattr(agent, "list_catalog", lambda db: [("S", "a.md", "A")])
    holder = AgentHolder()
    llm = FailingLLM(
        [
            ([], LLMError("down")),
            ([ToolCallPiece(id="call_1", name="ls", arguments={})], None),
            ([ToolCallPiece(id="call_2", name="ls", arguments={})], None),
            ([StreamPiece("content", "forced answer")], None),
        ]
    )
    with caplog.at_level(logging.INFO, logger="app.agent"):
        pieces = asyncio.run(_run(llm, holder, _settings(agent_max_rounds=2)))
    assert [type(p) for p in pieces] == [
        RetryPiece,
        ToolCallPiece,
        ToolCallPiece,
        StreamPiece,
    ]
    assert len(llm.requests) == 4  # 2 (round 1 + its retry) + 1 + the forced final
    assert llm.requests[3][1] is None  # the forced final, after round 2
    assert holder.tool_calls == 2
    msgs = [r.getMessage() for r in caplog.records]
    assert "agent tool=ls args={} round=1/2" in msgs
    assert "agent tool=ls args={} round=2/2" in msgs
    assert any("round cap reached (rounds=2)" in m for m in msgs)


# ---------- prompts: <tools> section (HIGH only) ----------
# NOTE (phase 70, task 02): these pins cover the phase-70 TOOLS_SECTION
# copy — the harness-aligned ls/read/grep names (the old phase-37/68
# names and the phase-37 per-tool budget line are gone; the round cap
# is the bound, not re-stated in the prompt, phase 45). The
# LOW/deflection path is untouched by this phase.


def test_high_prompt_carries_tools_section_after_documents() -> None:
    prompt = build_high_prompt([_doc("S", "a.md", "A", "A-CONTENT")])
    assert TOOLS_SECTION in prompt
    # Phase 70: the harness-aligned ls/read/grep copy.
    assert "`ls`" in prompt
    assert "`grep`" in prompt
    assert "`read`" in prompt
    assert "source: X | path: Y | title: Z" in prompt
    assert "locator, not a context-adder" in prompt
    assert "combined `source/path`" in prompt
    assert "Answer as soon as you have what you need" in prompt
    # The old names and the per-tool budget restatement are gone.
    for old in ("list_documents", "read_document", "search_documents"):
        assert old not in prompt
    assert "more than one extra document" not in prompt
    # After the mode body: <tools> follows </documents>.
    assert prompt.index("</documents>") < prompt.index("<tools>")
    assert prompt.rstrip().endswith("</tools>")


def test_high_prompt_tools_section_with_notes_and_kb() -> None:
    prompt = build_high_prompt(
        [_doc("S", "a.md", "A", "A-CONTENT")], notes=["be concise"], kb_overview="- KB"
    )
    assert prompt.index("<knowledge_base>") < prompt.index("<tuning>")
    assert prompt.index("<tuning>") < prompt.index("<documents>")
    assert prompt.index("<documents>") < prompt.index("<tools>")


def test_low_prompt_is_byte_identical_and_tool_free() -> None:
    # Phase 71: the LOW prompt carries the owner-permitted plain-text
    # line after the DEFLECT_MODE sentence (the marker-keying contract
    # is unchanged; the line must not leak into the HIGH prompt —
    # pinned in tests/unit/test_prompts.py).
    expected = (
        _base("LOW")
        + "\nDEFLECT_MODE: retrieval was weak — the titles below are the closest "
        "your notes come to the question. They are titles only; do not pretend "
        "they answer it. Use them to propose 2-3 alternative questions.\n"
        "Reply in plain text only — you have no tools in this mode.\n"
        + "- T1\n- T2"
    )
    assert build_deflect_prompt(["T1", "T2"]) == expected
    for prompt in (
        build_deflect_prompt(["T1"]),
        build_deflect_prompt(["T1"], notes=["be concise"]),
        build_deflect_prompt(["T1"], kb_overview="- KB"),
        build_deflect_prompt(["T1"], notes=["be concise"], kb_overview="- KB"),
    ):
        assert "<tools>" not in prompt
        assert TOOLS_SECTION not in prompt


# ---------- phase 71: the scaffolding recovery policy (deterministic only) ----------

#: The raw span from the 2026-09-03 incident (the E2E mock's trigger,
#: task 05) — a complete span the filter strips in full.
_INCIDENT_SPAN = (
    "<|tool_call_start|>[read(path='/homelab/backup-notes.md')]<|tool_call_end|>"
)


def test_correction_instruction_is_the_harness_constant() -> None:
    """Verbatim constant: the E2E mock (task 05) keys on a stable
    substring of it, so it must not drift."""
    assert agent.CORRECTION_INSTRUCTION == (
        "Your previous reply contained raw tool-call markup, which is not "
        "interpreted here. Answer the user's question directly in plain "
        "text — no tool syntax."
    )


def test_malformed_reply_error_subclasses_llm_error() -> None:
    assert issubclass(MalformedReplyError, LLMError)
    assert not issubclass(LLMError, MalformedReplyError)


def test_scaffolding_only_round_gets_exactly_one_recovery(
    monkeypatch: pytest.MonkeyPatch,
    caplog: pytest.LogCaptureFixture,
) -> None:
    """A round whose visible content is pure scaffolding → exactly TWO
    model requests: the normal round, then the ONE recovery —
    ``tools=None`` with :data:`CORRECTION_INSTRUCTION` folded into the
    original single system message (the user message stays last). The
    clean recovery answer ends the turn, the holder is untouched by the
    recovery, and the strip was captured in the warning log."""
    holder = AgentHolder()
    llm = ScriptedLLM(
        [StreamPiece("content", _INCIDENT_SPAN)],
        [StreamPiece("content", "The clean recovery answer.")],
    )
    with caplog.at_level(logging.WARNING, logger="app.agent"):
        pieces = asyncio.run(_run(llm, holder, _settings()))
    # Nothing of the round's scaffolding was yielded — only the recovery.
    assert pieces == [StreamPiece("content", "The clean recovery answer.")]
    assert len(llm.requests) == 2  # the round + the one recovery
    # The round: the original system prompt, the tools offered.
    assert llm.requests[0][1] == AGENT_TOOLS
    assert llm.requests[0][0] == [
        {"role": "system", "content": "SYSTEM_PROMPT"},
        {"role": "user", "content": "QUESTION"},
    ]
    # The recovery: no tools, a SINGLE system message carrying the folded
    # correction, the user message last.
    assert llm.requests[1][1] is None
    assert llm.requests[1][0] == [
        {
            "role": "system",
            "content": "SYSTEM_PROMPT\n" + agent.CORRECTION_INSTRUCTION,
        },
        {"role": "user", "content": "QUESTION"},
    ]
    # The recovery is a fixed policy, not a conversation: the holder is
    # untouched by it.
    assert holder.read_docs == [] and holder.tool_calls == 0
    # The turn total feeds the API layer's ``scaffold_stripped=N`` field.
    assert holder.scaffold_stripped == len(_INCIDENT_SPAN)
    # One strip warning per stripped span, the span truncated to 200 chars.
    strip_logs = [
        r
        for r in caplog.records
        if r.levelno == logging.WARNING and r.getMessage().startswith("agent: stripped")
    ]
    assert len(strip_logs) == 1
    message = strip_logs[0].getMessage()
    assert message.startswith(
        f"agent: stripped {len(_INCIDENT_SPAN)} chars of tool-scaffolding in round 1:"
    )
    assert _INCIDENT_SPAN[:200] in message


def test_scaffolding_twice_settles_with_malformed_reply(
    monkeypatch: pytest.MonkeyPatch,
) -> None:
    """The recovery answer is scaffolding again (a second empty reply) —
    terminal: :class:`MalformedReplyError` (an :class:`LLMError` subclass)
    after exactly two requests — no third request, no recovery of a
    recovery."""
    holder = AgentHolder()
    llm = ScriptedLLM(
        [StreamPiece("content", _INCIDENT_SPAN)],
        [StreamPiece("content", _INCIDENT_SPAN)],
    )
    with pytest.raises(MalformedReplyError) as excinfo:
        asyncio.run(_run(llm, holder, _settings()))
    assert isinstance(excinfo.value, LLMError)
    assert len(llm.requests) == 2  # exactly one recovery per turn
    assert llm.requests[1][1] is None
    assert agent.CORRECTION_INSTRUCTION in llm.requests[1][0][0]["content"]
    assert holder.scaffold_stripped == 2 * len(_INCIDENT_SPAN)


def test_scaffolding_with_real_content_needs_no_recovery(
    monkeypatch: pytest.MonkeyPatch,
) -> None:
    """A round with real visible content PLUS scaffolding: the clean
    content stands — one request only, the clean remainder yielded (no
    raw tokens), no correction in any system prompt."""
    holder = AgentHolder()
    llm = ScriptedLLM(
        [
            StreamPiece("content", "Here it is: "),
            StreamPiece("content", _INCIDENT_SPAN),
            StreamPiece("content", " hope that helps."),
        ],
    )
    pieces = asyncio.run(_run(llm, holder, _settings()))
    assert [p for p in pieces if isinstance(p, StreamPiece)] == [
        StreamPiece("content", "Here it is: "),
        StreamPiece("content", " hope that helps."),
    ]
    assert len(llm.requests) == 1  # no recovery
    for messages, _tools in llm.requests:
        assert all(agent.CORRECTION_INSTRUCTION not in m["content"] for m in messages)
    assert holder.scaffold_stripped == len(_INCIDENT_SPAN)


def test_clean_turn_carries_no_correction_and_no_strip(
    monkeypatch: pytest.MonkeyPatch,
) -> None:
    """A clean turn: one request, no correction in any system prompt,
    zero stripped (the log field stays 0 — uniform)."""
    holder = AgentHolder()
    llm = ScriptedLLM(
        [StreamPiece("thinking", "hmm "), StreamPiece("content", "a clean answer")]
    )
    pieces = asyncio.run(_run(llm, holder, _settings()))
    assert pieces == [
        StreamPiece("thinking", "hmm "),
        StreamPiece("content", "a clean answer"),
    ]
    assert len(llm.requests) == 1
    for messages, _tools in llm.requests:
        assert all(agent.CORRECTION_INSTRUCTION not in m["content"] for m in messages)
    assert holder.scaffold_stripped == 0


def test_empty_round_without_a_strip_keeps_today_behavior(
    monkeypatch: pytest.MonkeyPatch,
) -> None:
    """Round content 0 with NOTHING stripped (an empty/thinking-only
    answer) → return as today — no recovery, no error."""
    holder = AgentHolder()
    llm = ScriptedLLM([StreamPiece("thinking", "nothing to say, honestly")])
    pieces = asyncio.run(_run(llm, holder, _settings()))
    assert pieces == [StreamPiece("thinking", "nothing to say, honestly")]
    assert len(llm.requests) == 1
    assert holder.scaffold_stripped == 0


def test_scaffolding_round_with_tool_calls_needs_no_recovery(
    monkeypatch: pytest.MonkeyPatch,
) -> None:
    """A scaffolding-only round that ALSO carried tool calls: the tool
    ran, and the policy keys on the no-calls exit only — no recovery (the
    next round is a normal tools-offered round carrying the tool
    history)."""
    monkeypatch.setattr(agent, "list_catalog", lambda db: [("S", "a.md", "A")])
    holder = AgentHolder()
    llm = ScriptedLLM(
        [
            StreamPiece("content", _INCIDENT_SPAN),
            ToolCallPiece(id="call_1", name="ls", arguments={}),
        ],
        [StreamPiece("content", "the answer after the tool")],
    )
    pieces = asyncio.run(_run(llm, holder, _settings()))
    # The round's scaffolding was stripped (no raw delta), the tool frame
    # and the next round's answer flowed on.
    assert pieces == [
        ToolCallPiece(id="call_1", name="ls", arguments={}),
        StreamPiece("content", "the answer after the tool"),
    ]
    assert len(llm.requests) == 2
    assert llm.requests[1][1] == AGENT_TOOLS  # a normal round, not a recovery
    for messages, _tools in llm.requests:
        assert all(
            m["content"] is None or agent.CORRECTION_INSTRUCTION not in m["content"]
            for m in messages
        )
    assert holder.tool_calls == 1
    assert holder.scaffold_stripped == len(_INCIDENT_SPAN)


def test_recovery_after_tool_rounds_keeps_the_history(
    monkeypatch: pytest.MonkeyPatch,
) -> None:
    """A scaffolding-only answer round after a tool round: the recovery
    keeps the SINGLE (folded) system message at the front and the tool
    history intact behind it — no second system message, no duplicated
    correction."""
    monkeypatch.setattr(agent, "list_catalog", lambda db: [("S", "a.md", "A")])
    holder = AgentHolder()
    llm = ScriptedLLM(
        [ToolCallPiece(id="call_1", name="ls", arguments={})],
        [StreamPiece("content", _INCIDENT_SPAN)],
        [StreamPiece("content", "recovered after a tool round")],
    )
    pieces = asyncio.run(_run(llm, holder, _settings()))
    assert pieces == [
        ToolCallPiece(id="call_1", name="ls", arguments={}),
        StreamPiece("content", "recovered after a tool round"),
    ]
    assert len(llm.requests) == 3  # tool round + stripped round + recovery
    assert llm.requests[2][1] is None
    recovered = llm.requests[2][0]
    assert recovered[0] == {
        "role": "system",
        "content": "SYSTEM_PROMPT\n" + agent.CORRECTION_INSTRUCTION,
    }
    assert recovered[1] == {"role": "user", "content": "QUESTION"}
    assert len(recovered) == 4
    assert recovered[2]["role"] == "assistant"
    assert recovered[3] == {
        "role": "tool",
        "tool_call_id": "call_1",
        "content": "1 documents:\nsource: S | path: a.md | title: A",
    }
    assert sum(1 for m in recovered if m["role"] == "system") == 1
    assert holder.scaffold_stripped == len(_INCIDENT_SPAN)


def test_forced_final_scaffolding_only_settles_malformed(
    monkeypatch: pytest.MonkeyPatch,
) -> None:
    """The round-cap forced final (``tools=None``) is filtered too: a
    scaffolding-only forced answer never reaches the user raw — the turn
    settles with :class:`MalformedReplyError` (the same terminal
    semantics; this turn used no recovery, so nothing is doubled up)."""
    monkeypatch.setattr(agent, "list_catalog", lambda db: [("S", "a.md", "A")])
    holder = AgentHolder()
    llm = ScriptedLLM(
        [ToolCallPiece(id="call_1", name="ls", arguments={})],
        [StreamPiece("content", _INCIDENT_SPAN)],
    )
    with pytest.raises(MalformedReplyError):
        asyncio.run(_run(llm, holder, _settings(agent_max_rounds=1)))
    assert len(llm.requests) == 2
    assert llm.requests[1][1] is None  # the forced final — no recovery after it
    assert holder.scaffold_stripped == len(_INCIDENT_SPAN)